Download asSeq: A set of tools for the study of allele-specific RNA

asSeq: A set of tools for the study of allele-specific RNA-seq data
Wei Sun and Vasyl Zhabotynsky
February 22, 2013
1
Overview
> library(asSeq)
This vignette describes how to use R/asSeq to perform eQTL mapping using total expression and/or allelespecific expression, including a pipeline for input data preparation.
2
A brief Introduction
RNA-seq has become one of the major solutions for genome-wide inquiry of transcriptome variation. Allelespecific expression (ASE), which can be measured by RNA-seq but not by traditional microarray, provides
a new perspective of transcriptome variation. Allelic imbalance of gene expression may be due to cis-acting
genetic variant or parent-of-origin regulation. Currently this R package asSeq only provides support for
assessing cis-acting eQTL. Later we will include another set of functions that can dissect (cis-acting) genetic
effect and parent-of-origin effect. Here cis-acting regulation means a genetic variant on one allele (i.e., the
maternal or paternal allele) only affects the expression on the same allele. A cis-eQTL often refers to a localeQTL. In this document, we use a more precise definition of cis-eQTL such that they are the eQTL associated
with ASE. Interested readers are referred to Sun (2102) [1] and Sun and Hu (2012) [2] for more details. Sun
(2102) [1] provides the details of a statistical method for eQT mapping using both total expression and
ASE. Sun and Hu (2012) [2] is a review paper that reviews topics related with eQTL mapping using ASE or
isoform-specific expression. In addition to main functions for eQTL mapping using total expression or ASE,
asSeq also provides utility functions for quality control or extracting allele-specific RNA-seq reads, which
are the RNA-seq reads that overlap with heterozygous genetic markers. Figure 1 shows a complete pipeline
of eQTL mapping using ASE. Some additional scripts are provided for certain steps of this pipeline.
3
DNA Data Processing
We assume DNA genotype data are available, either from SNP array or DNA sequencing. It is possible to
call SNPs or indels in exonic regions from RNA-seq data, but we do not consider such approach here. First
we need to impute haplotype data in a larger number of SNPs, for example, the SNPs/indels from the 1000
Genome Project. Large number of SNPs/indels provide higher chance that a RNA-seq read overlaps with a
heterozygous genetic marker, hence is identified as allele-specific read. Haplotypes are necessary for eQTL
mapping and estimation of gene level ASE, which is defined by counting the total number of allele-specific
reads that may overlap with different SNPs/indels within a gene.
A set of R scripts are provided for haplotype phasing using MACH in folder scripts/MACH of this
R package. In these R scripts, we start with the SNP genotype from Affymetrix 6.0 array (less than
800,000 autosome SNPs), and impute the haplotypes on more than 36.8 million autosome SNPs from the
1000 Genome Project. Software MACH and all the 1000 Genome Project data needed for phasing can be
downloaded from http://www.sph.umich.edu/csg/abecasis/MACH/download/. Briefly, it takes four steps
to obtain haplotype estimates.
1. step1_prepare_for_MACH.R: prepare genotype data in merlin format as input files of MACH.
1
RNA Data Processing
DNA Data Processing
RAW RNA-seq Data of
FASTA/FASTAQ Format
Map Reads to
Reference Genome
DNA genotype data in
the study sample
haplotype data from
reference panels
Phasing, e.g., by
MACH or BEAGLE
Map Reads to
Individual-specific
haploid genomes
phased, imputed haplotype
data in the study sample
Mapped Reads
as BAM files
Counting
QC and counting
Remove reads with mapping ambiguity or low mapping quality
To count for allele-specific reads, remove those reads with low
mapping quality at the SNP position.
eQTL Mapping
TReC per gene per sample
ASE per gene per sample per allele
Covariates, such as
total reads per sample,
batch, gender, age, etc.
PCs estimated from
standardized TReC Data
eQTL mapping using
TReC and ASE
Figure 1: A pipeline for eQTL mapping using allele-specific RNA-seq data.
2. step2_split_ref.R: split reference haplotype files from 1000 Genome Project. We only use the data
of European Ancestor individuals for this example.
3. step3_run_MACH.R: run MACH using parallel computation for each split reference file.
4. step4_ligate.R: ligate the MACH results
4
RNA Data Processing
In most of the studies, RNA data processing mainly refer to RNA-seq mapping. The commonly used strategy
is to map the RNA-seq reads of all the individuals to the same reference genome. However, increasing evidence
has shown that an alternative approach, to map the RNA-seq reads of each individual to its own haploid
genomes, is more desirable in terms of identifying allele-specific reads. Certainly the two haploid genomes
of a diploid individual are not known in most of the cases. However they can be estimated (sometimes
referred to as pseudo-genome) by inserting phased indels and SNPs into the reference genome. These indels
and SNPs have to be phased and it is often safe to assume the phasing is accurate within a gene. Possible
switch errors of phasing in a longer range can be handled by specifically designed statistical methods in the
association testing step. This is a method that we are currently working on.
2
5
QC and Counting
5.1
Prepare the BAM files
The following codes provide some simple steps for BAM file preparation. Samtools is required for steps 1, 2,
and 4. In the 1st and 4th step, we count the number of reads in the original bam file and the processed bam
file. In the 2nd step, we sort the bam file by read names. In the 3rd step, we use an R function prepareBAM
of our R package asSeq to obtain uniquely mapped reads and apply a number of quality control criteria.
The function prepareBAM not only excludes those reads that fail QC, but also fixes the annotation in the
bam file, particularly the FLAG field of each record. For example, if one read is deleted, the FLAG filed of
its paired mate should be modified to reflect this change. prepareBAM can also do the sorting step, however
we found samtools can often perform the sorting step faster with smaller amount of memory.
# ---------------------------------------------------------------------------# 1. counting
# ---------------------------------------------------------------------------sami = "sample1"
bami = "sample1_lane1_xyz.bam"
ctF = sprintf("_count/count_%s.txt", sami)
cmd1 = sprintf("samtools view %s | wc -l >> %s\n", bami, ctF)
system(cmd1)
# ---------------------------------------------------------------------------# 2. sorting
# ---------------------------------------------------------------------------cmd2 = sprintf("samtools sort -n %s %s_sorted_by_name", bami, sami)
system(cmd2)
bamF = sprintf("%s_sorted_by_name.bam", sami)
# ---------------------------------------------------------------------------# 3. getUnique and filtering
# ---------------------------------------------------------------------------prepareBAM(bamF, sprintf("%s_sorted_by_name", sami), sortIt=FALSE)
# ---------------------------------------------------------------------------# 4. counting again
# ---------------------------------------------------------------------------cmd3 = sprintf("samtools view %s_sorted_by_name_uniq_filtered.bam | wc -l >> %s\n", sami, ctF)
system(cmd3)
5.2
Obtain the list of heterozygous SNPs of each individual (the “snpList” file)
Given haplotype data of each individual generated in the previous step, we will generate a list of heterozygous
SNPs to be used to identify allele-specific RNA-seq reads, which are the RNA-seq reads that overlap with
at least one heterozyogus SNP. We refer to this file as a “snpList” file. The first a few rows of a “snpList” file
look like the following:
chr1
chr1
chr1
chr1
1019668
1020428
1020496
1029889
G
T
A
C
A
C
G
T
The four columns are chromosome, position, allele 1 (the allele on haplotype 1) and allele 2 (the allele on
haplotype 2), respectively. The 3rd and 4th columns are essentially the two haplotypes of this individual. Which haplotype is the 1st/2nd haplotype of an individual can be arbitrarily defined. However, such
definition should remain the same thoroughout the analysis. This snpList file will be used by R function
3
extractAsReads to extract allele specific reads. Only the SNPs of heterozygous genotypes are needed in
this file, although the function extractAsReads can automatically filter out those SNPs with homozygous
genotypes, if there is any. An R code (get_snpList.R) that can generate such a snpList file can be in folder
scripts/snpList of this R package.
5.3
Extract the allele-specific reads
R function extractAsReads from our R package asSeq can be used to extract allele-specific sequence reads.
For example, the command can be
extractAsReads(input,
snpList, outTag="myOutput", prop.cut=.5, min.snpQ=10, phred=33)
Three main parameters are:
• input: input file name, this should be file of BAM format
• snpList: name of the file including a SNP list, must be tab-delimited file with four columns, chromosome, position, allele 1 and allele 2, without header
• outputTag: output files will be named as outputTag hap1.bam, outputTag hap2.bam, and outputTag hapN.bam.
There are several additional parameters, and three relatively more important ones are:
• prop.cut: one RNA-seq read may overlap with multiple heterozygous SNPs. We assign a read to one
of the two haplotypes if the proportion of those heterozygous SNPs suggesting the read is from that
haplotype is larger than prop.cut.
• min.snpQ: If the sequencing quality of the base-pair at a SNP is smaller than min.snpQ, we will not
use this SNP to call allele-specific read. The default value is 10, which means 10 above the baseline
phred score specified by parameter phred.
• phred: baseline phred score. Use 33 for Illumina 1.8+, and use 64 for sequences from previous Illumina
pipeline. see http://en.wikipedia.org/wiki/FASTQ_format for more details.
5.4
Counting the total number of reads and the number of allele-specific reads
We count the number of reads per exon set, which includes one or more exons. A single or pairedend read overlaps with an exon set if it overlaps with all the exons in this exon set and it does not
overlap with any other exon. In the existing annotations, the exons are often overlapping. We produce the annotation of non-overlapping exons for mouse (Mus_musculus.NCBIM37.67_data.zip) and human (Homo_sapiens.GRCh37.66_data.zip) in bed format, which can be downloaded from http://www.
bios.unc.edu/~weisun/software/isoform_files/. The scripts used to generate (and check) these nonoverlapping exons are included in this R package at folder scripts/getUniqExon.
The actual counting step can be done in different ways. We modify the intersectBed function of
bedtools to appropriately handle paired-end RNA-seq reads and implement it as an R function countReads
of another R package named isoform.
library(isoform)
countReads(bamFile, bedFile, outFile)
where bamFile is the name of the bam file after sorting by read name and after the QC steps, bedFile is
the name of the annotation file of non-overlapping exons, and outFile is the name of output file, which is
a text file that looks like the following:
17
3
2
11
chr10_100|ENSG00000107651|10;
chr10_100|ENSG00000107651|8;chr10_100|ENSG00000107651|10;
chr10_100|ENSG00000107651|8;chr10_100|ENSG00000107651|9;chr10_100|ENSG00000107651|10;
chr10_100|ENSG00000107651|9;chr10_100|ENSG00000107651|10;
4
Each line of this file is the count of of the number of RNA-seq reads overlapping a particular exon set. For
example, the 2nd line of the above example says that there are 3 RNA-seq reads that overlap with the exon
8 and 10 of gene cluster chr10 100, and these two exons are part of gene ENSG00000107651. A gene cluster
includes one or more genes that share at least one exon. This output file provides enough information for
the study of gene expression at isoform level. However, we will not study RNA isoform expression here. User
can easily collapse these counts at exon set level into gene level. If some exons are shared by two genes, we
will treat those exons shared by two genes as a gene.
Allele-specific expression has been extracted into two BAM files in the previous step and they can be
counted the same way.
6
eQTL mapping
6.1
Obtain genotype data for eQTL mapping
Denote the 1st and 2nd haplotype of an individual by h1 and h2 , respectively. As mentioned before, which
haplotype is the 1st/2nd haplotype of an individual can be arbitrarily defined. However, such definition
should be consistent with those used in the “snpList” file. For the k-th SNP, denote the alleles on haplotypes
h1 and h2 as h1k and h2k , respectively. We define a “phased” genotype for the k-th SNP as h1k h2k . Then if
a SNP has two alleles, A and B, the phased genotype can be AA, AB, AB or BB. In contrast to common
definition of genotype, AB and BA are distinct phased genotypes. The genotype data needed for eQTL
mapping should be a data matrix with each row for an individual and each column for a SNP. Each entry
in the matrix is 0, 1, 3, or 4, corresponding to genotype AA, AB, BA, or BB. Which allele of a SNP is A or
B can be defined arbitrarily. For example, we can consider the B allele as the minor allele.
6.2
eQTL mapping
Finally (yes, finally...), we obtain all the data needed for eQTL mapping using both total expression and
allele-specific expression by the R function trecase:
trecase(Y, Y1, Y2, X, Z, output.tag, p.cut)
These parameters/input data are explained in the following. We also explain which of the previous steps
generate such input data.
• Y: matrix of gene expression in terms of total read count (TReC). Each row corresponds to one sample
and each column corresponds to one gene. This data is generated at step 5.4.
• Y1/Y2:
column
haploid
snpList
matrices of allele-specific gene expression. Each row corresponds to one sample and each
corresponds to one gene. Specifically, Y1 and Y2 are the expression from first and second
genome, respectively. The first and second haploid genome is defined as the column order in
file, which is generated at step 5.2. Y1 and Y2 are generated at Step 5.4.
• X: matrix of confounding covariates. Each row corresponds to one sample and each column corresponds
to one variable. Usually these covariates include log(total number of reads per sample) and PCs
obtained from log transformed and total-number-of-reads-per-sample corrected expression matrix.
• Z: matrix of phased genotype data. Each row corresponds to one sample/individual and each column
corresponds to one SNP. Z must takes value of 0 (AA), 1 (AB), 3 (BA) or 4 (BB), where A and B
are two alleles of a SNP. AB and BA are different since they represent phased genotypes. Specifically,
AB means the 1st/2nd haplotype harbors the A/B allele, respectively, while BA means the 1st/2nd
haplotype harbors the B/A allele, respectively. This file is generated at step 6.1.
• output.tag The results of eQTL computation will be output into two files: output.tag eqtl.txt and
output.tag freq.txt. The former file includes (gene, SNP) pairs and their corresponding p-values, and
the latter file includes the distribution of all eQTL p-values, which may be useful to calculate FDR.
5
• p.cut p-value cut-off, only the eQTL associations with p-value smaller than p.cut are saved.
Joint model of both total expression and allele-specific expression will only be attempted if there are enough
allele-specific reads in enough samples, which are specified by parameters min.AS.reads and min.AS.sample.
We have implemented a likelihood ratio test (with p-value cutoff specified by transTestP) to assess whether
total read count and allele-specific read count give consistent association strength. If the answer is affirmative, joint model p-value will be reported as final p-value, otherwise the p-value from total read count will
be reported as the final p-value.
The columns of the output files include
• GeneRowID, MarkerRowID : row ID for gene and DNA marker, respectively.
• TReC_b, TReC_Chisq, TReC_df, TReC_Pvalue: regression coefficient, Chi-square test statistic, degree
of freedom, and p-value of TReC test. Similarly we have these columns for the results of ASE model
and Joint model.
• n_TReC, n_ASE: sample size for TReC and ASE model.
• n_ASE_Het: among those samples with ASE data, how many have heterozygous genotype on the DNA
marker.
• trans_Chisq, trans_Pvalue: test statistic and p-value to assess whether the eQTL is a trans-eQTL.
Small p-value implies trans-eQTL.
• final_Pvalue: final_Pvalue equals to TReC_Pvalue if it is a trans-eQTL, and it equals to Joint_Pvalue
otherwise.
There are two other functions for eQTL mapping using Total Read Count (TReC) only (function trec)
trec(Y, X, Z, output.tag, p.cut)
or allele-specific expression only (function ase).
ase(Y1, Y2, Z, output.tag, p.cut)
In most situations, these two functions are not necessary because function trecase already output the test
p-values using total read count only or ASE only. Function trec only uses total read count Y, but does
not need allele-specific read count Y1 and Y2. The genotype data for function trec (parameter Z) should
be numerical coding of genotype (e.g., additive coding with 0, 1, and 2 for AA, AB, and BB, respectively),
rather than phased genotype. Function ase only uses allele-specific reads Y1 and Y2 and the genotype data
Z should be the phased genotype, the same as the Z used by function trecase.
If one only keep the most significant eQTL for each gene (e.g., this is reasonable for a local eQTL scan),
an ideal solution of multiple testing correction is to evaluate the permutation p-value of the most significant
eQTL for each gene, and then choose a permutation p-value cutoff to control FDR across all the genes.
Permutation test is computationally intensive. Larger number of permutations is only needed for significant
associations. We implement this strategy in function trecaseP. One example of using this R function is
listed as below.
trecaseP(Y, Y1, Y2, X, Z, np.max=5000, np=c(20, 100, 500, 1000, 2500),
aim.p=c(0.5, 0.2, 0.1, 0.05, 0.02), confidence.p=0.01)
The parameters Y, Y1, Y2, X, Z are the same as those in function trecase. The extra parameters specify
how to do permutations. In this example, we will do at most 5000 permutations (np.max=5000). However,
we will terminate the permutation if at the number of permutations 20, 100, 500, 1000, and 2500 (np=c(20,
100, 500, 1000, 2500)), we already have >99% of confidence (confidence.p=0.01) that the permutation
p-value is larger than 0.5, 0.2, 0.1, 0.05, and 0.02 (aim.p=c(0.5, 0.2, 0.1, 0.05, 0.02)), respectively.
6
References
[1] Wei Sun (2012) A statistical framework for eQTL mapping using RNA-seq data. Biometrics. 68(1):1-11.
[2] Wei Sun and Yijuan Hu (2012), eQTL mapping using RNA-seq data, Statistics in Biosciences. in press
7